Abstract
Active inference is a domain-general theory of brain functioning which reconceptualises the perception-action interface in terms of a common process of minimization of sensory prediction errors. Such accounts have been extensively applied to the control of manual action guided by visual sensory feedback; however, they have received relatively little explicit attention in speech motor control. This is despite speech providing a critical test case, arguably being one of the most crucial and intricate of human sensorimotor functions. The application of active inference to speech motor control can allow crosspollination of decades of work from neighbouring disciplines, and could highlight where speech motor control mechanisms may be similar to, or differ from, those in other motor control domains, by establishing mechanistic explanation in common terms. We present here the first detailed description of an active inference framework of auditorily guided speech production. We compare the architecture of active inference models to existing computational models of speech motor control, and describe an active inference account of how compensation and adaptation result from perturbations of auditory feedback. We highlight several unique aspects of active inference, as well as emerging hypotheses for future empirical work. In particular, active inference accounts emphasise a role for proprioception in speech motor learning, and offer the potential to model the effects of other voices on speech production in phenomena such as phonetic convergence.